Hi 👋, I'm Natasha — a Quantitative Data Scientist working at the intersection of statistical modelling, machine learning, and complex systems.

My background spans business and economics, econometrics, risk analysis, and computer science. I started in business economics, moved into statistical modelling and uncertainty analysis during my PhD, and later added computer science to strengthen my engineering foundations.

I now build applied data and AI systems across the full workflow: data pipelines → modelling → evaluation → deployment.

I’m especially drawn to messy, high-stakes problems where the data is imperfect, the signal is partial, and the cost of getting it wrong is not trivial. That includes work in areas such as healthcare, risk, forecasting, trust, and decision support.

My work typically focuses on:

• statistical modelling, inference, and uncertainty analysis
• machine learning across structured, unstructured, image, and time-series data
• NLP and LLM-based systems where retrieval, evaluation, and reliability matter
• data pipelines, analytics engineering, and reproducible workflows in Python and SQL
• explainability, diagnostics, and decision support for real-world use

My approach is straightforward: strong statistical reasoning, practical engineering, and clear communication.

💼 Featured Projects

• Financial Conduct Risk Analyser

  Two-layer conduct risk platform screening 64 NZX and ASX listed companies using Altman Z-score, Beneish M-score, and Piotroski F-score, combined with LLM-based document analysis extracting structured risk signals from FMA enforcement notices, RBNZ reports, and UNCTAD regulatory documents. Includes an explicit validity assessment layer distinguishing screening instruments from decision instruments — with a documented construct validity failure illustrating where standard financial risk scores mislead in the NZ/AU context.

• Multi-Agent Risk Auditor

  Six specialist agents running in parallel across uploaded financial and regulatory documents, extracting conduct, financial, regulatory, operational, political, and environmental risk signals with citation tracking and human escalation. Forms a complete pipeline with the Financial Conduct Risk Analyser: quantitative screening identifies candidates; the auditor provides the evidence base for regulatory judgement.

• Financial Inclusion & Mobile Money Surge Forecasting

  End-to-end predictive analytics system forecasting mobile money adoption surges across 178 countries (2024–2028) using XGBoost on World Bank Findex and WDI data. Cross-validated ROC-AUC 0.858 ± 0.017. Opportunity Matrix dashboard designed for development finance capital allocation decisions — translating model outputs into country-level intelligence for non-technical stakeholders.

• Fraud Detection & Risk Scoring System

  Ensemble fraud detection system on 6.3M financial transactions combining XGBoost (ROC-AUC 0.9931) and Isolation Forest with SHAP-based explainability. Validated for robustness under severe distributional shift — testing whether model performance holds when the data environment changes adversarially.

• SDG Progress Analytics Dashboard

  End-to-end SDG monitoring and forecasting platform covering 268 countries and 28 indicators (2000–2023), integrating five institutional data sources (World Bank WDI, WHO, V-Dem, IMF, Global Forest Watch) into a unified panel. Per-country Holt-Winters projections to 2030 with 95% prediction intervals, traffic-light scorecard across 13 SDG goals, and live deployment on Hugging Face Spaces.

• Adaptive RL Policy Search for Histopathology Classification

  Reinforcement learning–guided augmentation policy search for breast cancer image classification using DenseNet201 on the BreakHis dataset. Demonstrates empirically that reward metric choice is a measurement validity problem: AUC-rewarded agents significantly outperform accuracy-rewarded agents on clinical recall. Corrects a systematic methodological flaw in published literature through patient-level stratified splitting.

• Spectral Drug Verification

  Verification workflow comparing measured spectral signatures against reference compound libraries, with focus on chemically similar classes, concentration differences, and classification reliability. Centred on signal quality, similarity structure, and decision confidence.

• Histopathology AI for Breast Cancer Detection

  Deep learning pipelines for breast cancer image classification using transfer learning, class-balancing strategies, and interpretability methods. Explores not just model performance but the practical challenges of medical image analysis in imbalanced settings.

• Air Quality Forecasting (PM10)

  Forecasting models for PM10 pollution in Auckland using regression and neural network approaches including LSTM. Focused on temporal modelling, feature design, and comparing methods for environmentally meaningful prediction.

• Big Data Sentiment Analysis of NASDAQ Companies

  Analysed more than 4 million tweets about NASDAQ-listed firms using Hadoop, MapReduce, Hive, and NLP techniques. Combines large-scale text processing with sentiment analysis to study public opinion in financial contexts.

• Customer Churn Prediction & Retention Analytics

  End-to-end churn modelling pipeline using XGBoost and SHAP with an interactive analytics layer for interpreting risk drivers and supporting retention decisions.

• Analytics Data Warehouse & ETL Design

  Structured analytics warehouse with ETL workflows and dimensional modelling principles to support reporting and downstream analysis.

• Histopathology AI for Breast Cancer Detection

  Developed deep learning pipelines for breast cancer image classification using transfer learning, class-balancing strategies, and interpretability methods. This work explores not just model performance, but the practical challenges of medical image analysis in imbalanced settings.

📝 How I Think & Where I Write

I write about systems under strain: data systems, decision systems, biological systems, and the human tendency to misunderstand all three.

My work sits somewhere between data science, AI, uncertainty, physiology, and philosophy of modern life. The themes vary, but the underlying interest is constant: complex systems, failure modes, and the gap between reality and the stories we tell about it.

I publish on Medium, Substack and LinkedIn. New connections are always welcome!

---

🛠 Tools & Stack

Languages

ML & Data Science

Engineering & Infrastructure

Visualisation & Dashboards